Add `git_oid_shorten` (unique OID minimzer)

Set of methods to find the minimal-length to uniquely identify every OID
in a list. Useful for GUI applications, commit logs and so on.

Includes stress test.

Signed-off-by: Vicent Marti <tanoku@gmail.com>

include/git2/oid.h

@@ -132,6 +132,60 @@ GIT_EXTERN(void) git_oid_cpy(git_oid *out, const git_oid *src);
  */
 GIT_EXTERN(int) git_oid_cmp(const git_oid *a, const git_oid *b);
 
+/**
+ * OID Shortener object
+ */
+typedef struct git_oid_shorten git_oid_shorten;
+
+/**
+ * Create a new OID shortener.
+ *
+ * The OID shortener is used to process a list of OIDs
+ * in text form and return the shortest length that would
+ * uniquely identify all of them.
+ *
+ * E.g. look at the result of `git log --abbrev`.
+ *
+ * @param min_length The minimal length for all identifiers,
+ *		which will be used even if shorter OIDs would still
+ *		be unique.
+ *	@return a `git_oid_shorten` instance, NULL if OOM
+ */
+git_oid_shorten *git_oid_shorten_new(size_t min_length);
+
+/**
+ * Add a new OID to set of shortened OIDs and calculate
+ * the minimal length to uniquely identify all the OIDs in
+ * the set.
+ *
+ * The OID is expected to be a 40-char hexadecimal string.
+ * The OID is owned by the user and will not be modified
+ * or freed.
+ *
+ * For performance reasons, there is a hard-limit of how many
+ * OIDs can be added to a single set (around ~22000, assuming
+ * a mostly randomized distribution), which should be enough
+ * for any kind of program, and keeps the algorithm fast and
+ * memory-efficient.
+ *
+ * Attempting to add more than those OIDs will result in a
+ * GIT_ENOMEM error
+ *
+ * @param os a `git_oid_shorten` instance
+ * @param text_oid an OID in text form
+ * @return the minimal length to uniquely identify all OIDs
+ *		added so far to the set; or an error code (<0) if an
+ *		error occurs.
+ */
+int git_oid_shorten_add(git_oid_shorten *os, const char *text_oid);
+
+/**
+ * Free an OID shortener instance
+ * 
+ * @param os a `git_oid_shorten` instance
+ */
+void git_oid_shorten_free(git_oid_shorten *os);
+
 /** @} */
 GIT_END_DECL
 #endif

src/oid.c

@@ -27,6 +27,7 @@
 #include "git2/oid.h"
 #include "repository.h"
 #include <string.h>
+#include <limits.h>
 
 static signed char from_hex[] = {
 -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, /* 00 */
@@ -166,3 +167,180 @@ int git_oid_cmp(const git_oid *a, const git_oid *b)
 {
 	return memcmp(a->id, b->id, sizeof(a->id));
 }
+
+
+typedef short node_index;
+
+typedef union {
+	const char *tail;
+	node_index children[16];
+} trie_node;
+
+struct git_oid_shorten {
+	trie_node *nodes;
+	size_t node_count, size;
+	int min_length, full;
+};
+
+static int resize_trie(git_oid_shorten *self, size_t new_size)
+{
+	self->nodes = realloc(self->nodes, new_size * sizeof(trie_node));
+	if (self->nodes == NULL)
+		return GIT_ENOMEM;
+
+	if (new_size > self->size) {
+		memset(&self->nodes[self->size], 0x0, (new_size - self->size) * sizeof(trie_node));
+	}
+
+	self->size = new_size;
+	return GIT_SUCCESS;
+}
+
+static trie_node *push_leaf(git_oid_shorten *os, node_index idx, int push_at, const char *oid)
+{
+	trie_node *node, *leaf;
+	node_index idx_leaf;
+
+	if (os->node_count >= os->size) {
+		if (resize_trie(os, os->size * 2) < GIT_SUCCESS)
+			return NULL;
+	}
+
+	idx_leaf = (node_index)os->node_count++;
+
+	if (os->node_count == SHRT_MAX)
+		os->full = 1;
+
+	node = &os->nodes[idx];
+	node->children[push_at] = -idx_leaf;
+
+	leaf = &os->nodes[idx_leaf];
+	leaf->tail = oid;
+
+	return node;
+}
+
+git_oid_shorten *git_oid_shorten_new(size_t min_length)
+{
+	git_oid_shorten *os;
+
+	os = git__malloc(sizeof(git_oid_shorten));
+	if (os == NULL)
+		return NULL;
+
+	memset(os, 0x0, sizeof(git_oid_shorten));
+
+	if (resize_trie(os, 16) < GIT_SUCCESS) {
+		free(os);
+		return NULL;
+	}
+
+	os->node_count = 1;
+	os->min_length = min_length;
+
+	return os;
+}
+
+void git_oid_shorten_free(git_oid_shorten *os)
+{
+	free(os->nodes);
+	free(os);
+}
+
+
+/*
+ * What wizardry is this?
+ *
+ * This is just a memory-optimized trie: basically a very fancy
+ * 16-ary tree, which is used to store the prefixes of the OID
+ * strings.
+ *
+ * Read more: http://en.wikipedia.org/wiki/Trie
+ *
+ * Magic that happens in this method:
+ *
+ *	- Each node in the trie is an union, so it can work both as
+ *	a normal node, or as a leaf.
+ *
+ *	- Each normal node points to 16 children (one for each possible
+ *	character in the oid). This is *not* stored in an array of
+ *	pointers, because in a 64-bit arch this would be sucking 
+ *	16*sizeof(void*) = 128 bytes of memory per node, which is fucking
+ *	insane. What we do is store Node Indexes, and use these indexes
+ *	to look up each node in the om->index array. These indexes are
+ *	signed shorts, so this limits the amount of unique OIDs that
+ *	fit in the structure to about 20000 (assuming a more or less uniform
+ *	distribution).
+ *
+ *	- All the nodes in om->index array are stored contiguously in
+ *	memory, and each of them is 32 bytes, so we fit 2x nodes per
+ *	cache line. Convenient for speed.
+ *
+ *	- To differentiate the leafs from the normal nodes, we store all
+ *	the indexes towards a leaf as a negative index (indexes to normal
+ *	nodes are positives). When we find that one of the children for
+ *	a node has a negative value, that means it's going to be a leaf.
+ *	This reduces the amount of indexes we have by two, but also reduces
+ *	the size of each node by 1-4 bytes (the amount we would need to
+ *	add a `is_leaf` field): this is good because it allows the nodes
+ *	to fit cleanly in cache lines.
+ *
+ *	- Once we reach an empty children, instead of continuing to insert
+ *	new nodes for each remaining character of the OID, we store a pointer
+ *	to the tail in the leaf; if the leaf is reached again, we turn it
+ *	into a normal node and use the tail to create a new leaf.
+ *
+ *	This is a pretty good balance between performance and memory usage.
+ */
+int git_oid_shorten_add(git_oid_shorten *os, const char *text_oid)
+{
+	int i, is_leaf;
+	node_index idx;
+
+	if (os->full)
+		return GIT_ENOMEM;
+
+	idx = 0;
+	is_leaf = 0;
+
+	for (i = 0; i < GIT_OID_HEXSZ; ++i) {
+		int c = from_hex[(int)text_oid[i]];
+		trie_node *node;
+
+		if (c == -1)
+			return GIT_ENOTOID;
+
+		node = &os->nodes[idx];
+
+		if (is_leaf) {
+			const char *tail;
+
+			tail = node->tail;
+			node->tail = NULL;
+
+			node = push_leaf(os, idx, from_hex[(int)tail[0]], &tail[1]);
+			if (node == NULL)
+				return GIT_ENOMEM;
+		}
+
+		if (node->children[c] == 0) {
+			if (push_leaf(os, idx, c, &text_oid[i + 1]) == NULL)
+				return GIT_ENOMEM;
+			break;
+		}
+
+		idx = node->children[c];
+		is_leaf = 0;
+
+		if (idx < 0) {
+			node->children[c] = idx = -idx;
+			is_leaf = 1;
+		}
+	}
+
+	if (++i > os->min_length)
+		os->min_length = i;
+
+	return os->min_length;
+}
+

tests/t01-rawobj.c

@@ -300,6 +300,96 @@ BEGIN_TEST(oid15, "convert raw oid to string (big)")
 	must_be_true(str && str == big && *(str+GIT_OID_HEXSZ+3) == 'Z');
 END_TEST
 
+
+BEGIN_TEST(oid16, "make sure the OID shortener doesn't choke on duplicate sha1s")
+
+	git_oid_shorten *os;
+	int min_len;
+
+	os = git_oid_shorten_new(0);
+	must_be_true(os != NULL);
+
+	git_oid_shorten_add(os, "22596363b3de40b06f981fb85d82312e8c0ed511");
+	git_oid_shorten_add(os, "ce08fe4884650f067bd5703b6a59a8b3b3c99a09");
+	git_oid_shorten_add(os, "16a0123456789abcdef4b775213c23a8bd74f5e0");
+	min_len = git_oid_shorten_add(os, "ce08fe4884650f067bd5703b6a59a8b3b3c99a09");
+
+	must_be_true(min_len == GIT_OID_HEXSZ + 1);
+
+	git_oid_shorten_free(os);
+END_TEST
+
+BEGIN_TEST(oid17, "stress test for the git_oid_shorten object")
+
+#define MAX_OIDS 1000
+
+	git_oid_shorten *os;
+	char *oids[MAX_OIDS];
+	char number_buffer[16];
+	git_oid oid;
+	size_t i, j;
+
+	int min_len = 0, found_collision;
+
+	os = git_oid_shorten_new(0);
+	must_be_true(os != NULL);
+
+	/*
+	 * Insert in the shortener 1000 unique SHA1 ids
+	 */
+	for (i = 0; i < MAX_OIDS; ++i) {
+		char *oid_text;
+
+		sprintf(number_buffer, "%u", (unsigned int)i);
+		git_hash_buf(&oid, number_buffer, strlen(number_buffer));
+
+		oid_text = git__malloc(GIT_OID_HEXSZ + 1);
+		git_oid_fmt(oid_text, &oid);
+		oid_text[GIT_OID_HEXSZ] = 0;
+
+		min_len = git_oid_shorten_add(os, oid_text);
+		must_be_true(min_len >= 0);
+
+		oids[i] = oid_text;
+	}
+
+	/*
+	 * Compare the first `min_char - 1` characters of each
+	 * SHA1 OID. If the minimizer worked, we should find at
+	 * least one collision
+	 */
+	found_collision = 0;
+	for (i = 0; i < MAX_OIDS; ++i) {
+		for (j = 0; j < MAX_OIDS; ++j) {
+			if (i != j && memcmp(oids[i], oids[j], min_len - 1) == 0)
+				found_collision = 1;
+		}
+	}
+	must_be_true(found_collision == 1);
+
+	/*
+	 * Compare the first `min_char` characters of each
+	 * SHA1 OID. If the minimizer worked, every single preffix
+	 * should be unique.
+	 */
+	found_collision = 0;
+	for (i = 0; i < MAX_OIDS; ++i) {
+		for (j = 0; j < MAX_OIDS; ++j) {
+			if (i != j && memcmp(oids[i], oids[j], min_len) == 0)
+				found_collision = 1;
+		}
+	}
+	must_be_true(found_collision == 0);
+
+	/* cleanup */
+	for (i = 0; i < MAX_OIDS; ++i)
+		free(oids[i]);
+
+	git_oid_shorten_free(os);
+
+#undef MAX_OIDS
+END_TEST
+
 static char *hello_id = "22596363b3de40b06f981fb85d82312e8c0ed511";
 static char *hello_text = "hello world\n";
 
@@ -518,6 +608,8 @@ BEGIN_SUITE(rawobjects)
 	ADD_TEST(oid13);
 	ADD_TEST(oid14);
 	ADD_TEST(oid15);
+	ADD_TEST(oid16);
+	ADD_TEST(oid17);
 
 	ADD_TEST(hash0);
 	ADD_TEST(hash1);